From theory to practice.
We believe that your company deserves products that are based on the latest and greatest scientific advances. Our suite of Interpolation Machines delivers security while not compromising on user experience, allowing your company to prioritise reliability and maximise efficiency in your AI solutions.
Spotlight research published by the founding team:
International Conference of Machine Learning (ICML) High Dimensional Learning Workshop.
Data Free Metrics are Not Reparameterisation Invariant
Properties of data-free metrics have been understood to highly correlate with the generalisation performance of neural networks. In this paper we show that data-free properties are not causal for generalisation and highlight the failure modes of weight based analysis of neural networks.
International Conference of Machine Learning (ICML) High Dimensional Learning Workshop.
Generalisation and Safety Critical Evaluations at Sharp Minima: A Geometric Reappraisal
Geometric properties of neural network have preferences to flat minima. We show how saftey critical evaluations can exist at sharp minima, identifying new interprectations of minima geometry.
International Conference of Machine Learning (ICML) Methods and Opportunities at Small Scale Workshop.
Decomposed Learning: An Avenue for Mitigating Grokking
Through empirical evaluations on the modular addition grokking task, we show that modified learning representations significantly reduce the effect of grokking and, in some cases, eliminates it.
The founding team have presented research at top conferences:
