Balancing conceptual depth with practical skill development is a persistent challenge in advanced machine learning (ML) education, where powerful frameworks can obscure underlying mathematical and computational principles. To address this, we define a new principled approach that we call full-stack machine learning (FSML), which emphasizes the construction of large language models and diffusion models from scratch. To evaluate the effectiveness of FSML, we conducted a classroom-based randomized controlled trial (N=208) in which FSML-based instruction was compared against a popular library-based instructional approach. We measured students' conceptual understanding through a specialized assessment and administered a survey capturing knowledge-gap awareness, curiosity, and cognitive load. We found that students who received FSML instruction performed approximately 10% better than control participants in a quiz on transformers and stable diffusion (p=0.006). They also showed increased curiosity and more positive affective responses, suggesting deeper engagement with ML fundamentals. Our findings indicate that our full-stack approach to ML education can improve student learning outcomes, potentially reshaping curricula for ML and other advanced computing topics.

Models that generate natural language explanations (NLEs) for their predictions have recently gained increasing interest.However, this approach usually demands large datasets of human-written NLEs for the ground-truth answers at training time, which can be expensive and potentially infeasible for some applications.When only a few NLEs are available (a fewshot setup), fine-tuning pre-trained language models (PLMs) in conjunction with promptbased learning has recently shown promising results.However, PLMs typically have billions of parameters, making full fine-tuning expensive.We propose SPARSEFIT, a sparse few-shot finetuning strategy that leverages discrete prompts to jointly generate predictions and NLEs.We experiment with SPARSEFIT on three sizes of the T5 language model and four datasets and compare it against existing state-of-the-art Parameter-Efficient Fine-Tuning (PEFT) techniques.We find that fine-tuning only 6.8% of the model parameters leads to competitive results for both the task performance and the quality of the generated NLEs compared to full finetuning of the model and produces better results on average than other PEFT methods in terms of predictive accuracy and NLE quality.

Neural network implementations are known to be vulnerable to physical attack vectors such as fault injection attacks. As of now, these attacks were only utilized during the inference phase. In this work, we explore a novel attack paradigm by injecting faults during the training phase in a way that the resulting network can be attacked during deployment without the necessity of further faulting. We discuss attacks against ReLU activation functions that make it possible to generate a family of malicious inputs, which are called fooling inputs, to be used at inference time to induce controlled misclassifications. Such malicious inputs are obtained by mathematically solving a system of linear equations that would cause a particular behaviour on the attacked activation functions, similar to the one induced in training through faulting. We call such attacks fooling backdoors as the faults at training phase inject backdoors into the network that allow an attacker to produce fooling inputs. We evaluate our approach against multi-layer perceptron networks and convolutional networks on a popular image classification task obtaining high attack success rates (60% - 100%) and high classification confidence when as little as 25 neurons are attacked while preserving high accuracy on the original classification task.

Identity theft is a major problem for credit lenders when there’s not enough data to corroborate a customer’s identity. Among super-apps—large digital platforms that encompass many different services—this problem is even more relevant; losing a client in one branch can often mean losing them in other services. In this paper, we review the effectiveness of a feature-level fusion of super-app customer information, mobile phone line data, and traditional credit risk variables for the early detection of identity theft credit card fraud. Through the proposed framework, we achieved better performance when using a model whose input is a fusion of alternative data and traditional credit bureau data, achieving a ROC AUC score of 0.81. We evaluate our approach over approximately 90,000 users from a credit lender’s digital platform database. The evaluation was performed using not only traditional ML metrics but the financial costs as well.