![Revisit consent button](../images/body-df29beacd7496220cc94.svg) We value your privacy We use cookies to enhance your browsing experience, serve personalised ads or content, and analyse our traffic. By clicking "Accept All", you consent to our use of cookies. Customise Reject All Accept All Customise Consent Preferences ![](../images/body-2159bb67f3de8ad870eb.svg) We use cookies to help you navigate efficiently and perform certain functions. You will find detailed information about all cookies under each consent category below. The cookies that are categorised as "Necessary" are stored on your browser as they are essential for enabling the basic functionalities of the site. ... Show more NecessaryAlways Active Necessary cookies are required to enable the basic features of this site, such as providing secure log-in or adjusting your consent preferences. These cookies do not store any personally identifiable data. * Cookie BCTempID * Duration 10 minutes * Description No description available. * Cookie __cf_bm * Duration 1 hour * Description This cookie, set by Cloudflare, is used to support Cloudflare Bot Management. * Cookie AWSALBCORS * Duration 7 days * Description Amazon Web Services set this cookie for load balancing. * Cookie _cfuvid * Duration session * Description Cloudflare sets this cookie to track users across sessions to optimize user experience by maintaining session consistency and providing personalized services * Cookie li_gc * Duration 6 months * Description Linkedin set this cookie for storing visitor's consent regarding using cookies for non-essential purposes. * Cookie __hssrc * Duration session * Description This cookie is set by Hubspot whenever it changes the session cookie. The __hssrc cookie set to 1 indicates that the user has restarted the browser, and if the cookie does not exist, it is assumed to be a new session. * Cookie __hssc * Duration 1 hour * Description HubSpot sets this cookie to keep track of sessions and to determine if HubSpot should increment the session number and timestamps in the __hstc cookie. * Cookie wpEmojiSettingsSupports * Duration session * Description WordPress sets this cookie when a user interacts with emojis on a WordPress site. It helps determine if the user's browser can display emojis properly. * Cookie BCSessionID * Duration 1 year 1 month 4 days * Description Blueconic sets this cookie as a unique identifier for the BlueConic profile. * Cookie _octo * Duration 1 year * Description No description available. * Cookie logged_in * Duration 1 year * Description No description available. * Cookie __Secure-YEC * Duration past * Description YouTube sets this cookie to stores the user's video player preferences using embedded YouTube video * Cookie __eoi * Duration 6 months * Description Description is currently not available. * Cookie AWSALBTGCORS * Duration 7 days * Description No description available. * Cookie login-status-p * Duration past * Description Description is currently not available. * Cookie AWSALBTG * Duration 7 days * Description No description available. * Cookie csrf_token * Duration session * Description No description available. * Cookie token_v2 * Duration 1 day * Description Description is currently not available. * Cookie D * Duration 1 year * Description Description is currently not available. * Cookie PHPSESSID * Duration session * Description This cookie is native to PHP applications. The cookie stores and identifies a user's unique session ID to manage user sessions on the website. The cookie is a session cookie and will be deleted when all the browser windows are closed. * Cookie VISITOR_PRIVACY_METADATA * Duration 6 months * Description YouTube sets this cookie to store the user's cookie consent state for the current domain. * Cookie cookietest * Duration session * Description The cookietest cookie is typically used to determine whether the user's browser accepts cookies, essential for website functionality and user experience. * Cookie __Host-airtable-session * Duration 1 year * Description This cookie is used to enable us to integrate the services of Airtable. * Cookie __Host-airtable-session.sig * Duration 1 year * Description This cookie is used to enable us to integrate the services of Airtable. * Cookie m * Duration 1 year 1 month 4 days * Description Stripe sets this cookie for fraud prevention purposes. It identifies the device used to access the website, allowing the website to be formatted accordingly. * Cookie BIGipServer* * Duration session * Description Marketo sets this cookie to collect information about the user's online activity and build a profile about their interests to provide advertisements relevant to the user. * Cookie __cfruid * Duration session * Description Cloudflare sets this cookie to identify trusted web traffic. * Cookie _GRECAPTCHA * Duration 6 months * Description Google Recaptcha service sets this cookie to identify bots to protect the website against malicious spam attacks. * Cookie __Secure-YNID * Duration 6 months * Description Google cookie used to protect user security and prevent fraud, especially during the login process. * Cookie cookieyes-consent * Duration 1 year * Description CookieYes sets this cookie to remember users' consent preferences so that their preferences are respected on subsequent visits to this site. It does not collect or store any personal information about the site visitors. Functional Functional cookies help perform certain functionalities like sharing the content of the website on social media platforms, collecting feedback, and other third-party features. * Cookie lidc * Duration 1 day * Description LinkedIn sets the lidc cookie to facilitate data center selection. * Cookie brw * Duration 1 year * Description No description available. * Cookie brwConsent * Duration 5 minutes * Description Description is currently not available. * Cookie WMF-Uniq * Duration 1 year * Description Description is currently not available. * Cookie loom_anon_comment * Duration 1 year * Description No description available. * Cookie loom_referral_video * Duration session * Description Description is currently not available. * Cookie VISITOR_INFO1_LIVE * Duration 6 months * Description A cookie set by YouTube to measure bandwidth that determines whether the user gets the new or old player interface. * Cookie yt-remote-connected-devices * Duration Never Expires * Description YouTube sets this cookie to store the user's video preferences using embedded YouTube videos. * Cookie ytidb::LAST_RESULT_ENTRY_KEY * Duration Never Expires * Description The cookie ytidb::LAST_RESULT_ENTRY_KEY is used by YouTube to store the last search result entry that was clicked by the user. This information is used to improve the user experience by providing more relevant search results in the future. * Cookie yt-remote-device-id * Duration Never Expires * Description YouTube sets this cookie to store the user's video preferences using embedded YouTube videos. * Cookie yt-remote-session-name * Duration session * Description The yt-remote-session-name cookie is used by YouTube to store the user's video player preferences using embedded YouTube video. * Cookie yt-remote-fast-check-period * Duration session * Description The yt-remote-fast-check-period cookie is used by YouTube to store the user's video player preferences for embedded YouTube videos. * Cookie yt-remote-session-app * Duration session * Description The yt-remote-session-app cookie is used by YouTube to store user preferences and information about the interface of the embedded YouTube video player. * Cookie yt-remote-cast-available * Duration session * Description The yt-remote-cast-available cookie is used to store the user's preferences regarding whether casting is available on their YouTube video player. * Cookie yt-remote-cast-installed * Duration session * Description The yt-remote-cast-installed cookie is used to store the user's video player preferences using embedded YouTube video. * Cookie cp_session * Duration 3 months * Description Codepen sets this cookie for Help systems found in the website. * Cookie loid * Duration 1 year 1 month 4 days * Description This cookie is set by the Reddit. The cookie enables the sharing of content from the website onto the social media platform. Analytics Analytical cookies are used to understand how visitors interact with the website. These cookies help provide information on metrics such as the number of visitors, bounce rate, traffic source, etc. * Cookie __hstc * Duration 6 months * Description Hubspot set this main cookie for tracking visitors. It contains the domain, initial timestamp (first visit), last timestamp (last visit), current timestamp (this visit), and session number (increments for each subsequent session). * Cookie hubspotutk * Duration 6 months * Description HubSpot sets this cookie to keep track of the visitors to the website. This cookie is passed to HubSpot on form submission and used when deduplicating contacts. * Cookie _ga * Duration 1 year 1 month 4 days * Description Google Analytics sets this cookie to calculate visitor, session and campaign data and track site usage for the site's analytics report. The cookie stores information anonymously and assigns a randomly generated number to recognise unique visitors. * Cookie _ga_* * Duration 1 year 1 month 4 days * Description Google Analytics sets this cookie to store and count page views. * Cookie __Host-psifi.analyticsTrace * Duration 6 hours * Description Description is currently not available. * Cookie __Host-psifi.analyticsTraceV2 * Duration 6 hours * Description Description is currently not available. * Cookie _gh_sess * Duration session * Description GitHub sets this cookie for temporary application and framework state between pages like what step the user is on in a multiple step form. * Cookie YSC * Duration session * Description YSC cookie is set by Youtube and is used to track the views of embedded videos on Youtube pages. * Cookie ajs_anonymous_id * Duration 1 year * Description This cookie is set by Segment to count the number of people who visit a certain site by tracking if they have visited before. * Cookie vuid * Duration 1 year 1 month 4 days * Description Vimeo installs this cookie to collect tracking information by setting a unique ID to embed videos on the website. Performance Performance cookies are used to understand and analyse the key performance indexes of the website which helps in delivering a better user experience for the visitors. * Cookie AWSALB * Duration 7 days * Description AWSALB is an application load balancer cookie set by Amazon Web Services to map the session to the target. * Cookie acq * Duration past * Description Description is currently not available. * Cookie acq.sig * Duration past * Description Description is currently not available. * Cookie ptc * Duration 2 years * Description No description available. Advertisement Advertisement cookies are used to provide visitors with customised advertisements based on the pages you visited previously and to analyse the effectiveness of the ad campaigns. * Cookie muc_ads * Duration 1 year 1 month 4 days * Description Twitter sets this cookie to collect user behaviour and interaction data to optimize the website. * Cookie guest_id_marketing * Duration 1 year 1 month 4 days * Description Twitter sets this cookie to identify and track the website visitor. * Cookie guest_id_ads * Duration 1 year 1 month 4 days * Description Twitter sets this cookie to identify and track the website visitor. * Cookie personalization_id * Duration 1 year 1 month 4 days * Description Twitter sets this cookie to integrate and share features for social media and also store information about how the user uses the website, for tracking and targeting. * Cookie guest_id * Duration 1 year 1 month 4 days * Description Twitter sets this cookie to identify and track the website visitor. It registers if a user is signed in to the Twitter platform and collects information about ad preferences. * Cookie bcookie * Duration 1 year * Description LinkedIn sets this cookie from LinkedIn share buttons and ad tags to recognize browser IDs. * Cookie __Secure-ROLLOUT_TOKEN * Duration 6 months * Description YouTube sets this cookie to manage feature rollout and experimentation. It helps Google control which new features or interface changes are shown to users as part of testing and staged rollouts, ensuring consistent experience for a given user during an experiment. * Cookie yt.innertube::nextId * Duration Never Expires * Description YouTube sets this cookie to register a unique ID to store data on what videos from YouTube the user has seen. * Cookie yt.innertube::requests * Duration Never Expires * Description YouTube sets this cookie to register a unique ID to store data on what videos from YouTube the user has seen. * Cookie session_tracker * Duration session * Description This cookie is set by the Reddit. This cookie is used to identify trusted web traffic. It also helps in adverstising on the website. * Cookie edgebucket * Duration session * Description Reddit sets this cookie to save the information about a log-on Reddit user, for the purpose of advertisement recommendations and updating the content. * Cookie did * Duration 1 year * Description Arbor sets this cookie to show targeted ads to site visitors.This cookie expires after 2 months or 1 year. Uncategorised Other uncategorised cookies are those that are being analysed and have not been classified into a category as yet. No cookies to display. Reject All Save My Preferences Accept All Toggle Search[](https://towardsdatascience.com/)Toggle Mobile Navigation [](https://towardsdatascience.com/) Publish AI, ML & data-science insights to a global community of data professionals. Sign in[Submit an Article](https://write.insightmediagroup.io) [Latest](https://towardsdatascience.com/latest/)[Editor's Picks](https://towardsdatascience.com/tag/editors-pick/)[Newsletter](https://towardsdatascience.com/tag/the-variable/)[ShipAI](https://towardsdatascience.com/shipai/) * * * [Write For TDS](https://towardsdatascience.com/submissions/) [LinkedIn](https://www.linkedin.com/company/towards-data-science/)[X](https://x.com/TDataScience)Toggle Search [Deep Learning](https://towardsdatascience.com/category/artificial-intelligence/deep-learning/) # Here Comes Mamba: The Selective State Space Model Part 3 - Towards Mamba State Space Models for Images, Videos and Time Series [Sascha Kirch](https://towardsdatascience.com/author/saschakirch/) September 3, 202418 min read ### [🐍 Towards Mamba State Space Models for Images, Videos and Time Series](https://towardsdatascience.com/tagged/mamba-image-video-signal) ![Image by Sascha Kirch.](../images/000.png) Image by [Sascha Kirch](https://medium.com/@SaschaKirch). > This is part 3 of my new multi-part series [🐍 Towards Mamba State Space Models for Images, Videos and Time Series](https://medium.com/@SaschaKirch/list/mamba-state-space-models-for-images-videos-and-timeseries-861ae0ad08fb). Mamba, the model to be said to replace the mighty Transformer, has come a long way from the initial idea of using state space models (SSMs) in deep learning. Mamba adds selectivity to state space models which results in Transformer-like performance while maintaining the sub-quadratic work complexity of SSMs. Their efficient selective scan is 40x faster than a standard implementation and can achieve 5x more throughput as a Transformer. Join me on this deep dive on Mamba where we will discover how selectivity solves the limitations of previous SSMs, how Mamba overcomes new obstacles that come with those changes and how we can incorporate Mamba into a modern deep learning architecture. > [**Mamba State Space Models for Images, Videos and Time Series**](https://medium.com/@SaschaKirch/list/861ae0ad08fb) * * * Learn this step by step with the interactive [Machine Learning](https://roadmap.sh/machine-learning) roadmap. ## Outline 1. [Introducing Selectivity to SSMs](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#0233) 2.1 [The Big Picture](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#76ae) 2.2 [What is Selectivity?](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#3f89) 2.3 [How Selectivity Affects Tensor Dimensions](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#7ca1) 2.4 [What's The Intuition Behind All Those Matrices?](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#77eb) 2. [Let's Talk About Matrix A, Again!](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#a616) 3.1 [Diagonal State Space Models](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#41c7) 3.2 [Mamba's Matrix A](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#16cc) 3. [Fast Training Without Convolutional Representation](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#cb62) 4.1 [Parallel Associative Scan](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#b1a8) 4.2 [Kernel Fusion](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#b23d) 4.3 [Recomputation of Gradients](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#2928) 4.4 [How Fast Is It?](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#34eb) 4. [Mamba as a General-Purpose Model](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#f98e) 5.1 [The Mamba Block](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#fac5) 5.2 [Where Are the Learnable Parameters?](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#1296) 5.3 [How Do We Initialize the Model?](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#6dd9) 5. [Wrapping Up Mamba](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#6d47) 6.1 [Summary](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#da74) 6.2 [Continue Reading Part 4](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#decf) * * * ## 1. Where We Left Off in Part 2 > [**Structured State Space Models Visually Explained**](https://towardsdatascience.com/structured-state-space-models-visually-explained-86cfe2757386) We ended Part 2 with the introduction of the S4 model. By constraining the state matrix _A_ to have a certain structure, the convolution kernel could be computed very efficiently making S4 30x faster using 400x less memory compared to previous SSMs, hence making it a feasible alternative to Transformers. We said that our structured state space models can be efficiently trained and efficiently inferred due to the duality of the recurrent and convolutional representations of LTI-Systems. ❗ **The issue though: An LTI-System treats all its inputs equally and therefore cannot make data dependent decisions.** * * * ## 2. Introducing Selectivity to SSMs Mamba addresses this problem by introducing a data dependent selection mechanism, but as we will see this comes at a cost. Later more about that. Let's start with some intuition. ### 2.1 The Big Picture A fundamental problem of sequence modelling is to decide to which extent the context (meaning previous inputs) is compressed into the state of the model. Remember that the **RNN** achieves fast inference due to its highly compressed state that models long-range dependencies ineffectively. A **Transformer** on the other hand, explicitly does not compress the context at all. Each token can attend to each other token at any time, allowing very effective modelling of long-range dependencies at the cost of quadratic work complexity _O(LΒ²)_. The **Structured SSM** approximates the context using the HiPPO matrix resulting in some compression, while it can be trained more efficiently as the RNN because of its convolutional representation. ![Figure 1: Efficiency vs. performance of the RNN, SSM, Transformer and the Mamba model. Image by Sascha Kirch.](../images/001.png) Figure 1: Efficiency vs. performance of the RNN, SSM, Transformer and the Mamba model. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). What we want is something in between, or let's better say the best of both worlds. A performant and efficient model. And spoiler alert: Mamba might be a promising candidate. ### 2.2 What is Selectivity? Instead of attending to the entire history like the Transformer and instead of compressing the entire history into a single state like RNNs and SSMs, we want to selectively compress just the relevant data to end up with a highly performant and efficient model. Efficient in this context means not only inference speed but also relates to other resources like memory and data needed to train the model. Let's look at this simply example of predicting my name: ![Figure 2: Selectively put focus on a subset of inputs. Image by Sascha Kirch.](../images/002.png) Figure 2: Selectively put focus on a subset of inputs. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). Up until now all matrices _A_ , _B_ and _C_ have been independent of the input data. They did not change for different time steps of the signal, hence we said we have a linear and time-invariant (LTI) system where the recurrent and convolutional representation were equal. ![Figure 3: Different representations of state space models. Image by Sascha Kirch.](../images/003.png) Figure 3: Different representations of state space models. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). Mamba changes this by making matrices _B_ and _C_ as well as the discretization parameter βˆ† time dependent. This means for every input token we end up with a different set of parameters! Recall that matrix _A_ is a structured matrix derived from the HiPPO matrix and approximates how the state evolves over time based on the history of previous states. Mamba keeps it that way. ### 2.3 How Selectivity Affects Tensor Dimensions Making the SSM selective introduces new dimension into the matrices. Explicitly, we now need to consider each token from a sequence of length _L_ parallelized to a batch size _B_. ![Figure 4: Tensor shapes of all matrices involved in the selective state space model. Image by Sascha Kirch.](../images/004.png) Figure 4: Tensor shapes of all matrices involved in the selective state space model. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). There are a few things to note here: 1. Matrix _A_ is of shape _DxN_ , and does not change for individual tokens. It is hence still time-invariant. We'll talk more about Matrix _A_ in [chapter 3](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#a616). 2. Matrix _A_ does perform a selection via the discretization step Ξ”, that is now dependent on the input. Hence the discretized matrix is of shape _BxLxDxN._ 3. Inside the SSM, matrices _B_ and _C_ are still vectors that are applied on input tokens or result in output tokens respectively, it's just that they are now stacked to build a tensor. 4. The discretized matrices are significantly larger than the non-discretized ones. More on that in [chapter 4](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#cb62). 5. These are the shapes of the matrices used in the SSM's equations, not the parameters that are learned! Those hide somewhere else. We will search for them in [chapter 5](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#f98e). **❓ But, what does all this even mean? What is the intuition behind all those matrices and why are only some selective?** ### 2.4 What's The Intuition Behind All Those Matrices? **Matrix A:** The state-to-state matrix decides what from the previous state propagates into the next state. **Matrix B:** the input-to-state matrix transforms any incoming token. It must be selective to filter out irrelevant data or to focus on important data to allow only relevant data to enter into the new state. It resembles some sort of gating mechanism. **Matrix C:** The state-to-output matrix transforms the state to an output token. It must be selective to decide which information from the state is required to materialize the output. Hence, like matrix _B_ , matrix _C_ has also some connections to a gating mechanism. **Matrix Ξ”:** The discretization parameter is often referred to as the step-size of converting the continuous input signal into a discrete sequence. While in a classical SSM, this might be intuitive, in a deep learning setup it is not, especially when working with already discretized data or other modalities like text tokens. So, let's look at the extreme cases: if _Ξ”β†’βˆž_ the system focuses more (or longer) on the input, hence in the state equation the state update is dominated by the contribution of the input and less by the previous state. If _Ξ”β†’0_ , the opposite happens. That means that Ξ” can be seen as some sort of weighting between the importance of new samples compared to the previous state. > [**Get an email whenever Sascha Kirch publishes πŸš€**](https://medium.com/@SaschaKirch/subscribe)[ _Get an email whenever Sascha Kirch publishes πŸš€ Looking to learn more about deep learning or simply stay up to dat_e...medium.com](https://medium.com/@SaschaKirch/subscribe) * * * ## 3. Let's Talk About Matrix A, Again! Again, we must talk about the state matrix _A_. Have you noticed something when looking at the dimension of matrix _A_? We were always talking about matrix _A_ being a square matrix of _NxN_ , but in Mamba its of shape _DxN_. What is happening? ### 3.1 Diagonal State Space Models Well, a lot has happened between the S4 paper and the release of Mamba. In particular, [diagonal state space (DSS)](https://arxiv.org/abs/2203.14343) models further simplified the DPLR (diagonal plus low-rank) formulation of S4 to be a simple diagonal matrix, by removing the low-rank term. When we talked about S4 in [part 2 of this series](https://medium.com/towards-data-science/structured-state-space-models-visually-explained-86cfe2757386?sk=479768bd75ecf8d410f902b7ad8c0836), we said having a diagonal matrix _A_ for the calculation of the convolution kernel _K_ would actually be the ideal case, since it reduces the number of computations drastically! ![Figure 5: Truly diagonal state matrix A as proposed by the DSS paper. Image by Sascha Kirch.](../images/005.png) Figure 5: Truly diagonal state matrix A as proposed by the DSS paper. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). In a [follow-up paper of S4's authors](https://arxiv.org/abs/2206.11893), they provided a thorough analysis of the effects on why this works and further suggests multiple new initialization schemes for the true diagonal case resulting in the S4D model. Long story short, Mamba uses the S4D-Real initialization, mentioned in a side note of the S4D paper, because it has the same spectrum as the original S4-LegS HiPPO Matrix. We will further talk about initialization in [chapter 5](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#f98e). ### 3.2 Mamba's Matrix A But now let's come back to Mamba's matrix _A_. Remember from [part 2 of the series](https://medium.com/towards-data-science/structured-state-space-models-visually-explained-86cfe2757386?sk=479768bd75ecf8d410f902b7ad8c0836) that if we have a diagonal matrix of _NxN_ , we only need to store _N_ values, meaning we can store it as a vector. The _DxN_ matrix of Mamba is really just _D_ times such a vector representing the diagonal matrix! ![Figure 6: Creating the diagonal matrices of the SSM from Matrix A. Image by Sascha Kirch.](../images/006.png) Figure 6: Creating the diagonal matrices of the SSM from Matrix A. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). This allows us to observe another interesting fact about Mamba and previous state space models: Each channel in _D_ is processed independently which can be interpreted as each channel gets its own SSM! This fact becomes important in [part 6](https://medium.com/towards-data-science/towards-mamba-state-space-models-for-images-videos-and-time-series-1e0bfdb5933a#efef) and [part 7](https://medium.com/towards-data-science/towards-mamba-state-space-models-for-images-videos-and-time-series-1e0bfdb5933a#1708) when we look into the MambaMixer and Mamba-2 respectively. * * * ## 4. Fast Training without Convolutional Representation Having different parameters for different timesteps implies that our model is now time-variant. Hence the convolution representation is not valid anymore and we loose the benefit of fast training. A big portion of the Mamba paper is about how to overcome this problem. They introduce three strategies resulting in their _**selective scan algorithm**_ : (1) parallel associative scan, (2) kernel fusion and (3) recomputation of gradients. > Note: These ideas are not completely new. The [S5 layer (Simplified State Space Layer for Sequence Modelling) ](https://arxiv.org/abs/2208.04933)applied the parallel scan and [H3 (Hungry Hungry Hippos)](https://arxiv.org/abs/2212.14052) already worked with kernel fusion to speed up the training. ### 4.1 Parallel Associative Scan Recall that the structured state space model is a time invariant system (LTI) since the matrices involved in updating the state (_A_ and _B_) are identical for each new input. That allowed either using a recurrent representation or a convolution representation. In Mamba this changes since the discretized matrices _A_ and _B_ are now dependent on the input. That implies that we cannot use the convolution representation to speed up the training and we are left with the recurrent representation, this time with different matrices for each new input and update of the state: ![Figure 7: Recurrent representation of a selective state space model. Image by Sascha Kirch.](../images/007.gif) Figure 7: Recurrent representation of a selective state space model. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). The recurrent execution has a work complexity of _O(L)_ , meaning doubling the length _L_ of the input sequence requires twice as much work to compute the final output. The one big issue with recurrency is that during training we don't want to calculate all the states sequentially, because we have access to the labels of all inputs beforehand. During inference, especially for an autoregressive sampling task (e.g. next word prediction) this would be totally fine though. ❓ **The question is, how can we speed up the sequential update of the state, where each state strongly depends on the previous one? It's not like that we could run it in parallel, right?** This particular type of recurrency, where each new state is the sum of the previous state and an input, is called _**scan**_. It is such a common problem in science and engineering, that even before we had parallel computers, in 1986 and 1990, techniques to parallelize a scan have been already been published. It wasn't until 2005 that these algorithms would be tested on a real GPU. ![Figure 8: Parallel scan algorithm. Image by Sascha Kirch.](../images/008.gif) Figure 8: Parallel scan algorithm. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). The parallel scan has a work complexity of _O(log L)_ , which is smaller than _O(L)_. The longer the input sequence _L_ , the higher the speed-up compared to the recurrent execution. Recall that a vanilla transformer has a work complexity of _O(LΒ²)_. ### 4.2 Kernel fusion As you might know, any computer has different types of memory. Generally speaking, the farther away from the processing unit, the larger the memory and the more time it takes to read the data (mainly due to properties of the PCB and other electronic components that do not allow higher frequencies). What usually happens in a GPU is that variables are read from the larger but farther away HBM (high bandwidth memory) into the faster SRAM to perform a calculation and then the result is written back into the HBM. If a calculation requires multiple steps, this process of reading, processing and writing is repeated until the entire calculation is done. If a calculation requires less time for the actual calculation as for reading and writing, it is called memory bounded aka. I/O bounded. ![Figure 9: GPU utilization vs. memory I/O. Image by Sascha Kirch.](../images/009.png) Figure 9: GPU utilization vs. memory I/O. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). I/O bound operations use their computing hardware quite inefficiently since they spend most of the time waiting for data in an idle state. Basically, you want to avoid reading large amounts of data too frequently. **❓ But what can we do? Or better said: what did the authors of Mamba do?** Instead of storing the discretized matrices _A_ and _B_ , only the continuous-time matrices are stored in global memory. Matrices _A_ , _B_ , _C_ and _Ξ”_ are read from the HBM into the SRAM and then the discretization is performed in SRAM. Here an example on how many values need to be read when either reading the continuous-time matrices or when reading the discretized matrices considering the shapes of the matrices as depicted earlier: ![Figure 10: Number of values needed to be read for discretized matrices vs. continuous matrices. Image by Sascha Kirch.](../images/010.png) Figure 10: Number of values needed to be read for discretized matrices vs. continuous matrices. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). Further, kernel fusion is applied to reduce the number of reads and writes from and to the HBM. In particular, they fuse the discretization, the selective scan and the multiplication with matrix _C_ into a single kernel. Hence, they read all the data needed for an entire forward pass through all states, process those data and write back the final output _y_ of shape _BxLxD_. The large states of shape _BxLxDxN_ are not materialized to avoid memory requirements and reading large amount of data. ![Figure 11: Impact of kernel fusion on the GPU utilization. Image by Sascha Kirch.](../images/011.png) Figure 11: Impact of kernel fusion on the GPU utilization. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). Having heard about kernel fusion and the different memories involved allows us to finally have a look on the iconic illustration from the Mamba paper showing the interaction of the different matrices to calculate the new states and the outputs: ![Figure 12: Selective State Space Model with Hardware-aware State Expansion. Image by A. Gu et.al., annotated by Sascha Kirch.](../images/012.png) Figure 12: Selective State Space Model with Hardware-aware State Expansion. Image by [A. Gu et.al.](https://arxiv.org/abs/2312.00752), annotated by [Sascha Kirch](https://medium.com/@SaschaKirch). Green indicates that those values are stored in the HBM while orange indicates that they are kept in the SRAM. Later in [chapter 5](https://towardsdatascience.com/here-comes-mamba-the-selective-state-space-model-435e5d17a451/#f98e) we will adapt this image to show what really happens and to find the trained parameters. > [**Paper Walkthroughs by Sascha Kirch**](https://medium.com/@SaschaKirch/list/89c7847da8e2) ### 4.3 Recomputation of Gradients Previously we saw that with kernel fusion we speed up inferring the model through a reduction of I/O operations and the fact that the hidden states are not saved in the HBM. However, to calculate the gradients and to back propagate the loss for the weight update, one needs those intermediate states. Since the kernel fusion is so effective, recomputing the intermediate states on demand as they are needed during back propagation is much more efficient than writing and reading the large intermediate states of shape _BxLxDxN_ from the HBM. ### 4.4 How Fast Is It? Having spent so much effort on making SSMs faster, one question remains: how fast is it compared to the most efficient attention implementation we know: [FlashAttention-2](https://arxiv.org/abs/2307.08691). ![Figure 13: Efficiency Benchmarks of Mamba. Image by A. Gu et.al., annotated by Sascha Kirch.](../images/013.png) Figure 13: Efficiency Benchmarks of Mamba. Image by [A. Gu et.al.](https://arxiv.org/abs/2312.00752), annotated by [Sascha Kirch](https://medium.com/@SaschaKirch). * * * ## 5. Mamba as a General-Purpose Model We are now equipped with the selective SSM, a state space model that is efficient, fast, scales linearly with the input sequence length _L_ and that is also selective since each input is processed with its own set of parameters. The only thing left for us is to do is to wrap it into a block and stack multiple blocks on top of each other to build a deep learning model; the same way as we would stack the multi-headed Transformer block in a Transformer-Decoder. ### 5.1 The Mamba Block Let's see what a model for an autoregressive next word prediction using Mamba could look like: > Note that the names of dimensions and names for the modules are the same as in the [official code implementation of the Mamba Block](https://github.com/state-spaces/mamba/blob/a07ff1b9ad2a4ac8b04eddf5eaaee5004f15aaf1/mamba_ssm/modules/mamba_simple.py#L31) to better understand how it can be configured and what is happening. ![Figure 14: Detailed block diagram of the Mamba-Block. In this example: L=4, d_model=3, expand=2, d_inner=6. Image by Sascha Kirch.](../images/014.png) Figure 14: Detailed block diagram of the Mamba-Block. In this example: L=4, d_model=3, expand=2, d_inner=6. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). The **input projection** simply projects each of the _L_ tokens into a `d_model`-dimensional space. The **head** does the prediction using one or more output tokens from the stack of **Mamba Blocks**. Inside the **Mamba Block** , we first add the `residuals` from the previous layer to the `hidden_states` and then output their sum as the next `residuals`. We normalize using RMS Layer Normalization and then linearly project from `d_model` dimensions into `2*d_inner` dimensions, where `d_inner` is defined as `expand*d_model`. Inside the code `expand` defaults to 2. This output of `in_proj` is then split into 2 different paths, both of which are now of dimension `[1, d_inner]` and both containing a SiLU non-linearity. The **Selective SSM block** is our state space model with the known state equations parametrized by the matrices _A_ , _B_ , _C_ and _Ξ”_. Note that the _vector dimension D_ from Figure 4 corresponds to `d_inner`. After the Selective SSM block, the two paths are combined, projected back from `d_inner` to `d_model` to form the final output. > Note that both linear projections are broadcasted over the _L_ tokens, meaning they use the same parameters for each of the tokens! ### 5.2 Where Are the Learnable Parameters? While inside the Mamba Block it is pretty clear where the learned parameters are located (i.e. input projection `in_proj`, output projection `out_proj` and the 1D convolution `conv1d`) it is still a mystery where the learned parameters hide inside the selective SSM Block. I think we can best observe what's going on if we adapt Figure 12, to show what really happens inside the selective SSM (single token and no batch axis): ![Figure 15: Selective State Space Model with Hardware-aware State Expansion including the discretized matrices and the learnable linear projections. Image by Sascha Kirch.](../images/015.png) Figure 15: Selective State Space Model with Hardware-aware State Expansion including the discretized matrices and the learnable linear projections. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). Again, a lot is going on this figure, so let's start with two observations: 1. Matrix _A_ has no subscript _t_ indicating it does not feature time-dependent selectivity, but the discretized Matrix _A_ does. 2. Matrices _A_ and _B_ are stored in the GPU's HBM, while the discretized matrices are only materialized inside the GPU SRAM as explained in section 4.2. Now let's have a look on what is happening. Note that the time-dependent matrices _B_ and _C_ are the result of the linear projection `x_proj` of the input _x_ at time _t._ Again, a single linear layer is used, likely for performance reasons. The output of `x_proj` is further projected from `dt_rank` to `d_inner` to form the final discretization matrix _Ξ”._ It is only Matrix _A_ we optimize directly (In reality we keep track of _log(A)_ for numerical stability during training). To conclude, inside the Selective SSM, we do not train the matrices used in the state equations, but we train the weights of linear projections that result in those matrices. It is only Matrix _A_ we are optimizing directly. And with that, there is only one question left to answer... ### 5.3 How Do We Initialize the Model? Not an easy question to be answered, since not only reading the papers is complicated due to the complicated matter of the subject, but also the code is quite convoluted. The first thing to notice: There is the initialization of a stand-alone Mamba Block and then there is a reinitialization, depending on the final architecture of the deep learning model that stacks multiple blocks together i.e. by using the `[_init_weights()](https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/models/mixer_seq_simple.py#L85-L115)` function. Let's fist look at the Mamba Block itself. All modules are initialized with pytorch's default initializer of the respective type of module (e.g. uniform for Conv1D and Linear) with two exceptions: Matrix _A_ and `dt_proj`, the linear projection used to construct _Ξ”._ Matrix _A_ is relatively straight forward to initialize using the S4D-Real initialization mentioned earlier: ![Figure 16: S4D-Real initialization of state matrix A. Image by Sascha Kirch.](../images/body-11ed54e23f17fd3445ca.png) Figure 16: S4D-Real initialization of state matrix A. Image by [Sascha Kirch](https://medium.com/@SaschaKirch). The timescale on the other hand is a bit more complicated. I recommend you to check out the ["How to Train your Hippo" paper](https://arxiv.org/abs/2206.12037) for more details. The short version is: The bias of `dt_proj` shall be initialized by log-uniformly sampling in a range of (_Ξ”min=0.001, Ξ”max= 0.1_). A very important side note in that paper that is easy to oversee: You can increase the performance if you set _1/Ξ”min_ to be approximately in the range of the sequence length. So 0.001 would be approximately 10Β³ samples. For example, look at this plot from the paper on a benchmark dataset with a sequence length of _L=16,384_ : ![Figure 17: Impact of choice of timescales on validation accuracy on Path-X benchmark. Image by A. Gu et.al., annotated by Sascha Kirch.](../images/body-552e2b566702034bf657.png) Figure 17: Impact of choice of timescales on validation accuracy on Path-X benchmark. Image by [A. Gu et.al.](https://arxiv.org/pdf/2206.12037), annotated by [Sascha Kirch](https://medium.com/@SaschaKirch). Finally, and without going too much into detail, here is the`[_init_weights()](https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/models/mixer_seq_simple.py#L85-L115)` function with some annotations that is used to overwrite already initialized weights when they are used inside the [MixerModel](https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/models/mixer_seq_simple.py#L118) and the [MambaLMHeadModel](https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/models/mixer_seq_simple.py#L215) the authors provided: > This function is actually called in both those classes meaning they overwrite the overwritten weights from the initial initialization _😡_ But jokes aside, if working under time pressure and finding the most performant solution there is not always enough time to write clean code. Yet another trade-off... ![Figure 18: Annotating the _init_weights\(\) function from the Mamba repository that is used to overwrite model weights after they have been initialized. Code taken from here, annotated by Sascha Kirch.](../images/body-8493461c18bb8eb648d8.png) Figure 18: Annotating the _init_weights() function from the Mamba repository that is used to overwrite model weights after they have been initialized. Code taken from [here](https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/models/mixer_seq_simple.py#L85-L115), annotated by [Sascha Kirch](https://medium.com/@SaschaKirch). * * * ## 6. Wrapping up Mamba ### 6.1 Summary With Mamba we now end up with a sequence model as powerful as the Transformer and as efficient as an RNN. We saw that while we lost the duality of the recurrent and the convolutional representation by adding selectivity, we were able speed up the model again by implementing the selective scan algorithm. The transformer is a general-purpose architecture, applied to all data modalities you can think of. So far, when talking about Mamba, we were only talking about sequence-like data such as text tokens or analog signals. **❓ But, can we apply Mamba to non-sequence data? Let's say images?** ### Thank You! > _🀝_ I'd like to t**hank you,** **t** he readers of my stories, for your constant support, clapping for my stories and, following me to never miss any of my latest articles. It is you who motivates me to continue writing, dive deep into complex topics, and become a better engineer/researcher by writing about them. You are awesome! ### 6.2 Continue Reading Part 4 > [**Vision Mamba: Like a Vision Transformer but Better**](https://towardsdatascience.com/vision-mamba-like-a-vision-transformer-but-better-3b2660c35848) In part 4 we will see how Mamba-like models can be applied on non-sequential data types such as images. We'll revisit certain design choices of the Vision Transformer (e.g. patch size, image resolution and position encodings) and see if they are still valid when using Mamba instead of attention. Vision Mamba ends up being 2.8x faster than [DeiT](https://arxiv.org/abs/2012.12877) and saves 86,8% GPU Memory, and we will see how this is possible. So stay tuned! * * * ## 7. Further Readings & Resources > [**Mamba State Space Models for Images, Videos and Time Series**](https://medium.com/@SaschaKirch/list/861ae0ad08fb) > [**Paper Walkthroughs by Sascha Kirch**](https://medium.com/@SaschaKirch/list/89c7847da8e2) ### Blog Posts * [Part 1]: [Towards Mamba State Space Models for Images, Videos and Time Series](https://towardsdatascience.com/towards-mamba-state-space-models-for-images-videos-and-time-series-1e0bfdb5933a?sk=8aaecd0fc979e1e95ac2a8e62946064b) * [Part 2]: [Structured State Space Models Visually Explained](https://towardsdatascience.com/structured-state-space-models-visually-explained-86cfe2757386?sk=479768bd75ecf8d410f902b7ad8c0836) * [A Visual Guide to Mamba and State Space Models](https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-mamba-and-state) ### Papers: * [Mamba: Linear-Time Sequence Modeling with Selective State Spaces](https://arxiv.org/abs/2312.00752) by A. Gu et.al., 1 Dec. 2023 * [H3: Hungry Hungry Hippos: Towards Language Modeling with State Space Models](https://arxiv.org/abs/2212.14052) by D. Fu et.al., 28 Dec. 2022 * [How to Train Your HiPPO: State Space Models with Generalized Orthogonal Basis Projections](https://arxiv.org/abs/2206.12037) by A. Gu et.al., 24 Jun. 2022 * [S4: Efficiently Modeling Long Sequences with Structured State Spaces](https://arxiv.org/abs/2111.00396) by A. Gu et.al., 31 Oct. 2021 * [S5: Simplified State Space Layers for Sequence Modeling](https://arxiv.org/abs/2208.04933) by J. Smith et. al., 9 Aug. 2022 * [HiPPO: Recurrent Memory with Optimal Polynomial Projections](https://arxiv.org/abs/2008.07669) by A. Gu et.al., 17 Aug. 2020 * [DSS: Diagonal State Spaces are as Effective as Structured State Spaces](https://arxiv.org/abs/2203.14343) by A. Gupta et. al., 27 Mar. 2022 * [S4D: On the Parameterization and Initialization of Diagonal State Space Models](https://arxiv.org/abs/2206.11893) by A. Gu et. al., 23 Jun. 2022 * [S4ND: Modeling Images and Videos as Multidimensional Signals Using State Spaces](https://arxiv.org/abs/2210.06583) by E. Nguyen et. al., 12 Oct. 2020 ### Code * [Mamba GitHub Repository](https://github.com/state-spaces/mamba) ## Register Now to Continue Reading ### Unlock Towards Data Science with a Free Account All articles on Towards Data Science are available to registered users. Create a free account today to access this article and thousands more. Register for freeSign in to your account * * * ![](../images/body-2648fe230469ec3ba2b1.png) Written By ## [Sascha Kirch](https://towardsdatascience.com/author/saschakirch/) [See all from Sascha Kirch](https://towardsdatascience.com/author/saschakirch/) [Deep Dives](https://towardsdatascience.com/tag/deep-dives/)[Deep Learning](https://towardsdatascience.com/tag/deep-learning/)[Mamba Image Video Signal](https://towardsdatascience.com/tag/mamba-image-video-signal/)[Parallel Computing](https://towardsdatascience.com/tag/parallel-computing/)[State Space Model](https://towardsdatascience.com/tag/state-space-model/) Share This Article * [](https://www.facebook.com/sharer/sharer.php?u=https%3A%2F%2Ftowardsdatascience.com%2Fhere-comes-mamba-the-selective-state-space-model-435e5d17a451%2F&title=Here%20Comes%20Mamba%3A%20The%20Selective%20State%20Space%20Model) * [](https://www.linkedin.com/shareArticle?mini=true&url=https%3A%2F%2Ftowardsdatascience.com%2Fhere-comes-mamba-the-selective-state-space-model-435e5d17a451%2F&title=Here%20Comes%20Mamba%3A%20The%20Selective%20State%20Space%20Model) * [](https://x.com/share?url=https%3A%2F%2Ftowardsdatascience.com%2Fhere-comes-mamba-the-selective-state-space-model-435e5d17a451%2F&text=Here%20Comes%20Mamba%3A%20The%20Selective%20State%20Space%20Model) Towards Data Science is a community publication. Submit your insights to reach our global audience and earn through the TDS Author Payment Program. [Write for TDS](https://towardsdatascience.com/submissions/) ## Related Articles * [![](../images/016.png)](https://towardsdatascience.com/zero-human-code-what-i-learned-from-forcing-ai-to-build-and-fix-its-own-code-for-27-straight-days/) ## [Zero Human Code: What I Learned from Forcing AI to Build (and Fix) Its Own Code for 27 Straight Days](https://towardsdatascience.com/zero-human-code-what-i-learned-from-forcing-ai-to-build-and-fix-its-own-code-for-27-straight-days/) [Artificial Intelligence](https://towardsdatascience.com/category/artificial-intelligence/) An experiment in pure AI development that challenges everything you've heard about coding assistants. [Daniel Bentes](https://towardsdatascience.com/author/danielbentes/) February 19, 202510 min read * [![](../images/017.jpg)](https://towardsdatascience.com/how-to-use-bert-from-the-hugging-face-transformer-library-d373a22b0209/) ## [How to use BERT from the Hugging Face transformer library](https://towardsdatascience.com/how-to-use-bert-from-the-hugging-face-transformer-library-d373a22b0209/) [Artificial Intelligence](https://towardsdatascience.com/category/artificial-intelligence/) How to use BERT from the Hugging Face transformer library for four important tasks [Saketh Kotamraju](https://towardsdatascience.com/author/saketh.kotamraju/) January 18, 202112 min read * [![](../images/018.png)](https://towardsdatascience.com/how-to-apply-artificial-bee-colony-algorithm-to-solve-unconventional-problems-a41c5098eb3a/) ## [How to Apply Artificial Bee Colony Algorithm to Solve Unconventional Problems](https://towardsdatascience.com/how-to-apply-artificial-bee-colony-algorithm-to-solve-unconventional-problems-a41c5098eb3a/) [Data Science](https://towardsdatascience.com/category/data-science/) Solve a puzzle for fun. Learn a new Algorithm. Win-win. Full code. [Naresh Ram](https://towardsdatascience.com/author/naresh-ram/) April 10, 202211 min read * [![](../images/019.png)](https://towardsdatascience.com/writing-custom-cuda-kernels-with-triton-abf6b6ad1168/) ## [Writing custom CUDA kernels with Triton](https://towardsdatascience.com/writing-custom-cuda-kernels-with-triton-abf6b6ad1168/) [Artificial Intelligence](https://towardsdatascience.com/category/artificial-intelligence/) Triton is a language compiler to write highly optimized CUDA kernels. Learn the basics of GPU programming and how to use Triton. [Kushajveer Singh](https://towardsdatascience.com/author/kushaj/) September 12, 20218 min read * [![](../images/020.png)](https://towardsdatascience.com/classifying-emotions-using-audio-recordings-and-python-434e748a95eb/) ## [Classifying emotions using audio recordings and Python](https://towardsdatascience.com/classifying-emotions-using-audio-recordings-and-python-434e748a95eb/) [Deep Learning](https://towardsdatascience.com/category/artificial-intelligence/deep-learning/) The theory of Sound and Acoustics could be used for cutting edge Emotion-Recognition technologies. here's how we can apply it using Python [Tal Baram](https://towardsdatascience.com/author/talbaram3192/) February 25, 202110 min read * [![](../images/021.jpg)](https://towardsdatascience.com/science-passion-and-the-future-of-multi-objective-optimization-5df55cd7327c/) ## [Science, Passion, and the Future of Multi-Objective Optimization](https://towardsdatascience.com/science-passion-and-the-future-of-multi-objective-optimization-5df55cd7327c/) [Science](https://towardsdatascience.com/category/society/science/) An In-Depth Interview with Professor Coello Coello [Emil Rijcken](https://towardsdatascience.com/author/emilrijcken/) October 6, 202313 min read * [![](../images/022.png)](https://towardsdatascience.com/paired-t-test-to-evaluate-machine-learning-classifiers-1f395a6c93fa/) ## [Paired t-test to evaluate Machine Learning classifiers using Python](https://towardsdatascience.com/paired-t-test-to-evaluate-machine-learning-classifiers-1f395a6c93fa/) [Artificial Intelligence](https://towardsdatascience.com/category/artificial-intelligence/) Learn how to compare classifiers using hypothesis testing [Marcello Politi](https://towardsdatascience.com/author/marcellopoliti/) July 6, 20225 min read * [![](../images/023.jpg)](https://towardsdatascience.com/getting-tensorflow-developer-certified-3975d1a26050/) ## [Getting TensorFlow Developer certified](https://towardsdatascience.com/getting-tensorflow-developer-certified-3975d1a26050/) [Deep Learning](https://towardsdatascience.com/category/artificial-intelligence/deep-learning/) Steps you can use to pass, too - even if you got an M1 Mac [Pascal Janetzky](https://towardsdatascience.com/author/pascaljanetzky/) July 31, 20217 min read * [![](../images/024.jpg)](https://towardsdatascience.com/what-is-batch-normalization-46058b4f583/) ## [What is batch normalization?](https://towardsdatascience.com/what-is-batch-normalization-46058b4f583/) [Artificial Intelligence](https://towardsdatascience.com/category/artificial-intelligence/) How does it help? [NVS Yashwanth](https://towardsdatascience.com/author/nvsyashwanth/) September 18, 20204 min read * [YouTube](https://www.youtube.com/c/TowardsDataScience) * [X](https://x.com/TDataScience) * [LinkedIn](https://www.linkedin.com/company/towards-data-science/) * [Threads](https://www.threads.net/@towardsdatascience) * [Bluesky](https://bsky.app/profile/towardsdatascience.com) [](https://towardsdatascience.com/) Your home for data science and AI. The world's leading publication for data science, data analytics, data engineering, machine learning, and artificial intelligence professionals. Subscribe to Our Newsletter[Write For TDS](https://towardsdatascience.com/submissions/)β€’[About](https://towardsdatascience.com/about-towards-data-science-d691af11cc2f/)β€’[Advertise](https://contact.towardsdatascience.com/advertise-with-towards-data-science)β€’[Privacy Policy](https://towardsdatascience.com/privacy-policy/)β€’[Terms of Use](https://towardsdatascience.com/website-terms-of-use/) Β© Insight Media Group, LLC 2026β€’[The New Stack](https://thenewstack.io)β€’[roadmap.sh](https://roadmap.sh)β€’[Creator Network](https://insightmediagroup.io/creator-network) word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word word mmMwWLliI0fiflO&1 mmMwWLliI0fiflO&1 mmMwWLliI0fiflO&1 mmMwWLliI0fiflO&1 mmMwWLliI0fiflO&1 mmMwWLliI0fiflO&1 mmMwWLliI0fiflO&1 #### Recommended Articles Close * [ ![](https://h030.towardsdatascience.com/rest/v2/contentStores/d7d3a4ab-7caf-4383-bbd2-fbf7fb4f1787/items/3642/image?etag=1780830904480&width=500&height=500) Check Your Biases ](https://towardsdatascience.com/check-your-biases-ba2685a7f799/ "Check Your Biases") * [ ![](https://h030.towardsdatascience.com/rest/v2/contentStores/d7d3a4ab-7caf-4383-bbd2-fbf7fb4f1787/items/607930/image?etag=1782124566054&width=500&height=500) The Subset Sum Problem Solved in Linear Time for Dense Enough Inputs ](https://towardsdatascience.com/subset-sum-problem-solved-in-linear-time-for-dense-enough-inputs/ "The Subset Sum Problem Solved in Linear Time for Dense Enough Inputs") * [ ![](https://h030.towardsdatascience.com/rest/v2/contentStores/d7d3a4ab-7caf-4383-bbd2-fbf7fb4f1787/items/3660/image?etag=1780176246070&width=500&height=500) Does Your Company Have a Data Strategy? ](https://towardsdatascience.com/does-your-company-have-a-data-strategy-e8944dbec4ae/ "Does Your Company Have a Data Strategy?") * [ ![](https://h030.towardsdatascience.com/rest/v2/contentStores/d7d3a4ab-7caf-4383-bbd2-fbf7fb4f1787/items/4973/image?etag=1782232567105&width=500&height=500) Machine Learning Tricks, Learning with ChatGPT, and Other Recent Must-Reads ](https://towardsdatascience.com/machine-learning-tricks-learning-with-chatgpt-and-other-recent-must-reads-f277648a1921/ "Machine Learning Tricks, Learning with ChatGPT, and Other Recent Must-Reads")