Differentiable Swift enables the world’s fastest AI training and unlocks the world of autonomous systems. By building differentiable computing into our technology at the compiler level, PassiveLogic is investing in the fundamental technologies to enable generative autonomy.
Don’t believe the graph? Then watch our 2023 launch event to see all three AI frameworks run some physics code head-to-head, in real-time. PassiveLogic blows them out of the water through the power of differentiability. Differentiable Swift runs 322x faster than Google’s TensorFlow and 238x faster than PyTorch.
We’ve run a benchmark comparison of a building thermal model, optimized via gradient descent and implemented in differentiable Swift vs. PyTorch vs. TensorFlow. The results are clear: differentiable Swift proves to be the fastest solution, and demonstrates the opportunities created by PassiveLogic’s investment in this language feature. See the repo for instructions on how to perform the benchmarks on your own.
Differentiability is a type of reversible computing that allows code to not only run forward but also backward. This enormously powerful technique is the magic behind extremely fast gradient descent and deep learning. At PassiveLogic, we’re unlocking the potential of generalized differentiability for any code you can dream up — not just conventional neural networks — and making the first differentiable systems language for all programmers and their use cases.
To power the leap to generative autonomous systems, we need an incredibly advanced AI. We use Swift because it offers first-class general purpose differentiability built right into the language, runtime composability, and no boundary between your production code and your machine learning code. With fast runtime differentiation, you can both train and perform inference at the edge — enabling run-time solutions like multi-dimensional inferencing that solves neural networks in any direction.
PassiveLogic has put significant work into addressing issues in differentiable Swift to meet the needs of specific applications, such as the requirements for autonomous systems. Learn more about PassiveLogic’s advancements, and why we’ve built the world’s largest differentiable computing team to support this development.
Differentiable Swift began as an experimental feature in the Swift for TensorFlow project at Google. Our head of Compiler Brad Larson was a member of that original team and continues to develop the technology here at PassiveLogic. Read the original manifesto to learn about the motivations for this technology, from multiple expert contributors in this domain.
PassiveLogic’s work in differentiable Swift opens up deep learning to solve a new set of problems beyond its traditional limits. It powers control pathfinding and real-time decision-making at the edge, enabling truly autonomous systems. It also powers generative design to solve challenges such as optimized energy modeling or finding the best sensor layout for a building.
Gradient descent is a technique used to navigate graph-based state spaces and find absolute minimum points: those minimum points often represent optimized states. Gradient descent, when combined with ultra-fast differentiable Swift, allows us to cover really complex graphs and find the optimal solution fast, taking the guesswork out of deep learning and eliminating costly training times.
IntrospectionKit is a principled set of tools and APIs within our Differentiable Swift framework that leverages graph compute to enable abstraction from any data structure. At PassiveLogic, it connects differentiable Swift with Quantum: the digital twin data standard for autonomous systems. With IntrospectionKit systems become self-aware as they gain the ability to write their own queries anywhere within Quantum.
InferenceKit allows PassiveLogic to understand more with our models than what can be directly observed. To have a complete digital twin, we can’t measure everything that’s happening in our buildings — the AI must fill in the gaps . InferenceKit achieves this, through differentiable Swift, by starting with known values and then inferring other values along the graph model. Now you can have full visibility into your assets even where there are no sensors.