Mini-Course on Reliable and Efficient AI

University of Pavia — May 25–29, 2026

Lecturer: Osvaldo Simeone (Northeastern University London)

Overview

This mini-course starts by exploring two frontier directions in the design of AI systems that are computationally efficient: neuromorphic computing and quantum machine learning. The course concludes with a rigorous treatment of calibration methods that provide formal statistical guarantees on the behaviour of AI models in practice. The material is drawn from the lecturer's recent research and monographs, offering a unified perspective that bridges theory and emerging applications.

Slides overview

Part I — Modern Neuromorphic Computing 

The rapid growth of AI has brought unprecedented capabilities but also escalating energy demands. This first part of the course examines neuromorphic computing as a principled response to that challenge, revisiting brain-inspired design ideas in the light of modern deep learning architectures.

Drawing on the paper Modern Neuromorphic AI: From Intra-Token to Inter-Token Processing (arXiv:2601.00245), the lectures explore how concepts such as discrete and sparse activations, recurrent dynamics, and non-linear feedback — long central to spiking neural networks (SNNs) — reappear in contemporary architectures including state-space models and transformers. A key organising lens is the distinction between intra-token processing (transformations across the channels of a single input vector, as in image classification with SNNs) and inter-token processing (transformations across a sequence of vectors, as in language modelling). The lectures trace the connections between these regimes and show how neuromorphic principles illuminate the design space of modern AI.

The second part of this block turns to communication systems as a compelling application domain for neuromorphic AI. Based on the paper Neuromorphic Wireless Cognition: Event-Driven Semantic Communications for Remote Inference (arXiv:2206.06047, Chen, Skatchkovsky & Simeone, IEEE Transactions on Cognitive Communications and Networking, 2023) and follow-up works, the lectures present an end-to-end design for neuromorphic wireless IoT systems. In this framework, spike-based sensors, SNNs, and impulse-radio transceivers are jointly designed so that energy is consumed only when meaningful events occur, yielding substantial improvements in time-to-accuracy and energy efficiency relative to conventional frame-based digital approaches.

Topics covered:

Slides Part I

Part II — Quantum Statistics and Machine Learning

The second part of the course introduces quantum machine learning (QML) from an engineer's perspective, with an emphasis on statistical foundations and learning-theoretic analysis.

Lectures followin part the monograph An Introduction to Quantum Machine Learning for Engineers (arXiv:2205.09510), published in Foundations and Trends in Signal Processing, and the textbook Classical and Quantum Information Theory, published by Cambridge University Press. The material is self-contained, starting from the mathematical description of quantum states, operations, and measurements before building up to parametrised quantum circuits (PQCs) — the dominant programming model for gate-based quantum computers in the current noisy intermediate-scale quantum (NISQ) era. PQCs are shown to be capable of addressing combinatorial optimisation problems, implementing generative models, and performing classification and regression.

A particular focus is placed on fundamental statistical tasks in the quantum setting: binary hypothesis testing and classification, quantum state discrimination, and the problem of learning from data when measurements are destructive and irreversible. The lectures examine how classical notions of sample complexity, generalisation, and model capacity translate — and sometimes break down — in the quantum regime, where the copy complexity arising from the no-cloning theorem introduces qualitatively new challenges for learning algorithms.

Topics covered:

Slides Part II

Part III — Calibration for Reliable AI 

The final part of the course addresses one of the most pressing practical challenges in deploying AI: ensuring that model outputs come with rigorous, interpretable reliability guarantees.

Based on the recent survey Conformal Calibration: Ensuring the Reliability of Black-Box AI in Wireless Systems (arXiv:2504.09310, Simeone, Park & Zecchin, 2025), the lectures review the framework of conformal calibration — a collection of computationally lightweight, statistically rigorous tools that operate on top of any pre-trained model without requiring retraining or fine-tuning. The standard train-and-deploy paradigm, which treats AI models as best-effort black boxes, is shown to be inadequate for safety-critical or high-stakes applications. Conformal calibration addresses this gap through two complementary mechanisms: pre-deployment calibration, which uses held-out data to produce prediction sets or select hyperparameters with provable coverage guarantees; and online monitoring, which detects and mitigates failures as they occur during deployment.

Although the survey is framed around wireless network applications, the techniques are general and apply broadly to any AI system where reliability and uncertainty quantification matter.

Topics covered:

Slides Part III

Prerequisites

A background in probability, linear algebra, and basic machine learning is assumed. No prior knowledge of quantum mechanics or neuromorphic hardware is required; all necessary concepts will be developed from first principles during the course.

Reference Materials

Part IV (bonus) — Measures of evidence

The fourth part of the course develops a rigorous, modern treatment of statistical hypothesis testing, with an emphasis on the principles that underlie both classical and emerging approaches. The central question is how to quantify evidence against a null hypothesis in a way that is statistically valid, interpretable, and computationally tractable. The lectures examine this question systematically, starting from the construction of p-values, progressing through group-invariance-based tests, and culminating in the theory of e-values and anytime-valid sequential inference.

The first block covers the foundations of p-value construction. After distinguishing aleatoric uncertainty, epistemic uncertainty, and statistical evidence, the lectures introduce p-values as measures of evidence and critically examine their strengths and limitations. A general recipe for constructing valid p-values under composite nulls is presented, illustrated through several canonical examples: rank p-values for the exchangeability null, sign-flip p-values for the symmetry null, Hoeffding p-values for bounded means, and permutation p-values under group-invariance nulls. A unifying theme is the tradeoff between the generality of the null — which governs the tractability of the test and the strength of the conclusion — and its specificity, which determines statistical power and data efficiency.

The second block introduces e-values as an alternative and often more powerful currency for expressing statistical evidence. An e-value is a non-negative statistic whose expectation under the null is at most one; its reciprocal is always a valid p-value by Markov's inequality, but e-values carry strictly richer structure. The lectures establish key properties: post-hoc validity (the rejection level can be chosen after observing the data, without inflating Type I error), multiplicative combination across independent experiments, and robustness to optional stopping. The connection to sequential inference is made precise through e-processes and Ville's inequality, which extends the classical Markov bound from fixed sample sizes to the supremum over all stopping times. Betting-based constructions of e-processes for composite nulls — including Kelly-optimal strategies for testing bounded means — are derived and analysed.

Topics covered:

Slides Part IV

Prerequisites

A background in probability is assumed. 

Reference Materials