Novartis will establish an “AI innovation lab” designed to enable its research and commercial employees to use Microsoft’s artificial intelligence solutions across operations, with the aim of transforming how new treatments are discovered, developed, and commercialized.
The pharma giant has selected Microsoft as its partner for a multi-year AI-based R&D effort, through a collaboration whose value was not disclosed. The companies did say, however, that their overall investment in AI would include project funding, subject-matter experts, technology, and tools.
The companies laid out two objectives for the AI lab. One is creating new AI models and applications that will allow every Novartis associate to access the company’s datasets combined with Microsoft’s AI solutions.
The other objective is using the power of AI to tackle some of the hardest computational challenges within life sciences—including generative chemistry, image segmentation, and analysis for smart and personalized delivery of therapies, as well as optimization of cell and gene therapies at scale.
The companies agreed to carry out joint research activities that will include co-working environments on Novartis’ headquarters campus in Basel, Switzerland; at Novartis Global Service Center in Dublin; and at Microsoft Research Lab in the U.K.—starting with tackling personalized therapies for macular degeneration; cell & gene therapy; and drug design.
Microsoft and Novartis also agreed to partner in supporting future programs across its two objectives by developing and applying next-generation AI platforms and processes.
“Reimagine medicine”
“As Novartis continues evolving into a focused medicines company powered by advanced therapy platforms and data science, alliances like this will help us deliver on our purpose to reimagine medicine to improve and extend patients’ lives,” Novartis CEO Vas Narasimhan said in a statement. “Pairing our deep knowledge of human biology and medicine with Microsoft’s leading expertise in AI could transform the way we discover and develop medicines for the world.”
Novartis said the partnership with Microsoft will help it fulfill its “strategic pillar” of “going big on data and digital” in order to be a leading developer of new treatments. To that end, Novartis said, it focuses its efforts around four strategic digital priority areas:
- Scaling 12 digital lighthouse projects: Build a strong foundation and jumpstart our digital transformation
- Make Novartis digital: Sharing, learning, and talent acquisition
- Become the number-one partner in the tech ecosystem: Bridge Novartis with external expertise
- Bolder moves: Lead through future disruptive healthcare scenarios with large-scale partnerships.
Added Microsoft CEO, Satya Nadella: “Our strategic alliance will combine Novartis’ life sciences expertise with the power of Azure [Microsoft’s cloud computing platform] and Microsoft AI. Together, we aim to address some of the biggest challenges facing the life sciences industry today and bring AI capabilities to every Novartis employee so they can unlock new insights as they work to discover new medicines and reduce patient costs.”
One way of breaking open the artificial neural network piñata and getting at all the candy inside is to understand ReLU as a switch.
With an electrical switch n volts in gives n volts out when on, a 45 degree line when graphed. I’m not going to explain off.
The dot product (weighted sum) of a number of dot products is still a dot product.
A particular input to a ReLU network causes the switches to each be in a particular state, on or off. Then there is a particular linear projection from the input to the output.
Since ReLU switches at zero there are no sudden discontinuities in the output with gradual changes in the input.
For each output neuron and a particular input you can calculate an equivalent single weighted sum. You can see what the weighted sum is looking at in the input and calculate metrics like the angle between weight vector and the input vector.
If the angle is small the weighted sum is providing some error correction via central limit theorem type averaging.
If the angle is close to 90 degrees and the neuron output large then the weight vector length must be large and magnify strongly noise at the inputs to the weighted sum.
The FFT and for example the Walsh Hadamard transform are also linear algebra devices. Directly incorporating them in switched artificial neural networks is entirely valid as they too have dot product behavior. Then you have the possibility of computing a network layer in nlog(n) operations compared to the conventional burden of n*n operations. Google fixed filter bank neural networks if you can accept the unconventional.