👉🕳👈 Closing the Gap Between Deep Learning Software and Hardware

📝 Editorial 

The topic of artificial intelligence(AI)-first hardware continues to gain prominence within the deep learning community. These days, it proves to be challenging to write high-performance deep learning models without optimizing for the underlying infrastructure. Specific TensorFlow or PyTorch programs can look very different when executing in different hardware architectures, such as Google TPUs or NVIDIA Jetson chips. This problem becomes worse as it's highly fragmented across the increasing number of deep learning frameworks and the new ecosystem of specialized deep learning hardware. In general, deep learning seems to have brought back the times in which software development needed to be concerned with the underlying hardware, and those dependencies are increasing rather than decreasing.  

Decoupling the dependencies between deep learning software and hardware is one of the existential challenges of the next decade of AI. The good news is that there are some very exciting projects trying to address this challenge. Just this week, AIOps startup OctoML raised a $28M series B to tackle this specific challenge. OctoML builds on top of the open-source  Apache TVM machine learning compiler framework, which was designed to bridge the gap between deep learning frameworks and specialized hardware and accelerators. Efforts like OctoML/TVM are often ignored, but they are nothing but essential to continue making deep learning accessible to mainstream developers.


🔺🔻TheSequence Scope – our Sunday edition with the industry’s development overview – is free. To receive high-quality content about the most relevant developments in the ML world every Tuesday and Thursday, please subscribe to TheSequence Edge 🔺🔻

🗓 Next week in TheSequence Edge:

Edge#73: the concept of Meta-Learning as a form of AutoML; OpenAI’s Reptile model for efficient meta-learning; the Auto-Keras framework.  

Edge#74 is about how Uber, Google, DeepMind and Microsoft Train Models at Scale.


Now, let’s review the most important developments in the AI industry this week

🔎 ML Research

Massively Parallel Graph Computations

Google Research published a paper proposing Adaptive Massively Parallel Computation (AMPC), a model that addresses some of the limitations of Map Reduce to allow the execution of graph algorithms at large scale ->read more on Google Research blog

AutoML for Time Series

IBM Research published a paper detailing the research behind Watson’s AutoAI Time Series ->read more on IBM Research blog

Transformers for Video Intelligence

Facebook AI Research (FAIR) published a paper unveiling TimeSformer, a transformer-based model for video analysis ->read more Facebook AI blog


🤖 Cool AI Tech Releases

Horovod + Ray 

Uber published some insightful details about an architecture based on Horovod and Ray to run machine learning models at scale ->read more on Uber blog

Follow us on Twitter


💸 Money in AI

Begin Capital invests from day zero to Series A and is looking for new pirates, winners, and robots to join their portfolio. Find out more about them at www.begincl.com. Statistically, 0.015% of website visitors get funding!* Today, experts from Begin Capital commented on a few of last week investment rounds*:

  • Shield AI raised $90 million to accelerate its work on defense AI technologies. The startup aims to develop best-in-class AI capabilities for conflict zones. Shield AI believes autonomous systems will bolster overall safety in military-active regions. 

Begin Capital says: The US Government is concerned with the growing pace of military AI development in China. We anticipate more startups working on military AI to win government contracts and raise large venture rounds in the coming years as seen per Anduril and Skydio mega-rounds.

  • AI-enabled SaaS platform Optibus raised a $107 million Series C round. Optibus platform provides analytics for people and vehicles moving directions, navigation, scheduling, rostering, and other guidance.

Begin Capital says: We see more cities around the globe adopting AI. Smart cities emergence bolster investment attractiveness of urban mobility and planning platforms as seen per $900 million Intel x Moovit acquisition and $100 million Via x Remix deal. 

  • No-code AI solution for enterprises Noogata raised a $12 million seed round. The company created domain-dedicated libraries of building blocks that target the most common use cases where AI can have its greatest impact such as e-commerce, retail, customer service and others.

Begin Capital says: Historically low-code DevOps market was ahead of low-code data analytics with the example of Weights & Biases $45 million round as stated in one of the previous TheSequence Scopes. However, with Noogata’s round or the UK-based startup Actable AI (just raised first pre-seed round) we see that this field is growing fast and becomes more attractive for VC funds to make their bet on.

  • Tomi AI raised their first $1 million round to enhance digital-marketing effectiveness with the help of AI. They integrate their models with first-party data and connect to ad platform APIs to explicitly boost digital campaign’s ROI.

Begin Capital says: While many MarTech companies try to propose simply better UX for marketers – solutions for efficiency (easiness & time) rather than effectiveness (ROI and value-add to ad platforms), companies like Tomi AI are the new wave. These startups harness new technologies and enable ML to impact efficiency of ad platforms in industries with low conversion rates and high customer acquisition cost.

Leave a comment

*sponsored

Other interesting rounds:

  • ML startup OctoML raised a $28 million Series B funding round. Built on Apache TVM, OctoML aims to accelerate model performance while enabling seamless deployment of models across any hardware platform, cloud provider, or edge device.

  • AI patent intelligence company PatSnap raised $300 million in series E funding. The company leverages computer vision, NLP and other AI technology to provide their clients with competitive intelligence as well as patent insights needed to take their products from ideation to commercialization.

  • Identity startup Socure raised $100 million in a Series D funding round. Its predictive analytics platform applies AI and ML techniques with trusted online/offline data intelligence from email, phone, address, IP, device, velocity, and the broader internet to verify identities in real-time. 

  • Global location intelligence startup SafeGraph raised $45 million in a Series B. It uses AI to create and maintain mobility datasets. SafeGraph customers gain access to not only location information, but spatial hierarchy metadata and place traffic data.

  • Network-centric AI startup Torch.AI raised $30 million. Using ML models for optical character recognition, natural language processing, sentiment analysis, and more, Torch.AI’s Nexus platform connects disparate apps, systems, cloud services, and databases to enable high-speed data reconciliation and high-speed processing.

  • AI-powered knowledge process automation startup DeepSee.ai raised a $22.6 million Series A funding round. Instead of widespread ‘robotic process automation’, the startup offers what they call ‘knowledge process automation’ (KPA). They built KPA to mine unstructured data, operationalize AI-powered insights, and automate results into real-time action for the enterprise. 

  • AI startup Sherpa raised an additional $8.5 million to their Series A round. Based on proprietary AI algorithms with conversational and predictive abilities, its digital assistant technology learns about the users and anticipates their needs before they ask.