📲 Why Mobile Deep Learning is Tougher Than You Think
📝 Editorial
Mobile devices represent a primary runtime for our daily interactions with machine learning models. However, the vast majority of machine learning experiences in mobile devices are delivered in a server-side architecture, with the machine learning model executing in a cloud environment and exposing results to mobile apps via an API. From training, personalization to computational resource consumption, the mobile deep learning paradigm presents many inefficiencies for mobile architectures. The holy grail of mobile deep learning is to build models that can execute natively and efficiently in mobile devices. These days, we have mobile deep learning frameworks in popular deep learning stacks like PyTorch or TensorFlow that easily allow you to adapt deep learning models to mobile architectures.
However, don’t get too excited yet. Mobile deep learning is one of those things that looks cool and simple from the outside and becomes a royal nightmare when you get into it.
The challenges with building deep learning models that execute on mobile devices are many and highly diverse. The heterogeneity of hardware architecture makes it extremely hard to optimize models for different mobile runtimes. Regularly training models that need to execute in millions of devices requires creative techniques such as federated learning, which is still not fully battle-tested. Finally, the developer experience from mobile deep learning is very limited compared to frameworks or platforms for models that execute in server-side topologies. However, the research and technology in the space is advancing rapidly. Just this week, there were releases from TensorFlow and ONNX as well as some relevant research papers that were tackling different challenges of mobile deep learning systems. These releases keep signaling the importance of mobile deep learning for some of the top tech companies in the world. For now, let’s try to not underestimate the many challenges of mobile deep learning systems and approach those problems with a dose of excitement and healthy skepticism.
🔺🔻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#97: Policy optimization RL Methods (PO RL); how Google trained RL agents to master the most popular sport in the world; DeepMind’s BSuite is a unique Benchmark System for RL models.
Edge#98: how OpenAI built RL agents that mastered Montezuma’s Revenge by going backwards.
Now, let’s review the most important developments in the AI industry this week
🔎 ML Research
Addressing Toxicity in GPT-3
OpenAI published a research paper and a sample dataset detailing a technique used to improve the behavioral values of GPT-3 using a small dataset ->read more on OpenAI blog
Semantic Understanding in Text-Based Games
Microsoft Research published two papers detailing techniques to advance semantic understanding in reinforcement learning using text-based games ->read more on Microsoft Research blog
Representation Learning in Small Devices
Google Research published a paper introducing FRILL, a representation learning model that can compute large feature sets using minimum resource consumptions and can be adapted to different on-device architectures ->read more on Google Research blog
🤖 Cool AI Tech Releases
NetHack Challenge
Facebook AI Research(FAIR) launched the NetHack Challenge, a game-based challenge to advance research in some of the most important areas of reinforcement learning ->read more on FAIR blog
TensorFlow PluggableDevice
TensorFlow open sourced PluggableDevice, an architecture for registering different hardware runtimes without requiring changes in the machine learning model ->read more on TensorFlow blog
ONNX 1.8
Microsoft, Facebook and others unveiled the new version of the ONNX Runtime for machine learning interoperability. The new framework includes acceleration for mobile and web architectures as well as an improved experienced for PyTorch developers->read more on Microsoft ML blog
💬 Useful tweet
Oh, wow.
💸 Money in AI
The U.K. government and IBM announced a five-year $297.5 million partnership to accelerate discovery and innovation with AI and quantum computing. The joint STFC – IBM program looks to add 60 new scientists, interns, and students.
🦄 AI talent management startup Eightfold AI raised $220 million in a Series E round led by SoftBank Vision Fund 2. The startup uses deep learning and artificial intelligence to help companies find, recruit and retain workers. Its current valuation is $2.1 billion.
🦄 AI-powered transcription startup Verbit raised a $157 million Series D round led by Sapphire Ventures. Verbit combines top-notch automatic speech recognition technology with fact-checking by skilled professional transcribers, which guarantees the highest accuracy. The current valuation is more than $1 billion.
Digital adoption solutions platform Whatfix raised $90 million in a funding round led by SoftBank Vision Fund 2. The startup provides chatbot-style guidance on how to use apps, creating a personalized journey for each user.
People analytics startup ChartHop raised $35 million in a Series B funding round led by Andreessen Horowitz. The startup brings disparate sources of people data together in a dynamic, visual, & actionable platform.
Dataflow automation startup Prefect raised $32 million in a Series B funding round led by Tiger Global. Its platform reduces engineering time by automating pipelines and workflows of different complexity.
Buyer intelligence startup Slintel raised $20 million in a Series A led by GGV. Slintel uncovers various forms of buying intelligence to make the process of identifying high-intent prospects more intuitive and fully data-driven.
Clinical data marketplace Mendel raised $18 million in a Series A round led by DCM. The company uses AI tools to analyze clinical data, including medical history and genetic analysis, from cancer patients.
AI-powered accounting platform Osome raised $16 million in a Series A funding round. It is a digital business assistant that offers online accounting services for small and medium businesses, automating the most routine tasks.
AI-driven freelance management startup Stoke raised $15.5 million in a Series A round of funding. The company built a freelance management system (FMS) to manage independent contractors and guarantee compliance through continuous analysis.
Working ethics and compliance platformVault Platform raised$8.2 million in a Series A led by Gradient Ventures. The startup leverages AI for misconduct detection and reporting.