CV / Sensing ML Fundamentals Oral

Sensing / computer-vision engineer phone screen mixing a quick coding warm-up with rapid-fire ML fundamentals: CNN vs MLP, how ResNet works, activation-function choices, and basic probability. Requirements Manager-led phone screen for a sensing / computer-vision engineer role. After a short coding warm-up, the interviewer drills core ML fundamentals out loud, including: CNN vs MLP: when and why a convolutional network is preferred over a plain multilayer perceptron. ResNet: what residual connections are and why they help train deep networks. Activation functions: whether something like y = 0.01x can serve as an activation layer (a leaky-style linear function), and what that implies about non-linearity. Probability / statistics basics such as the Gaussian distribution. Notes The coding warm-up here was a quick string-reversal task (write it recursively, then re-implement with an explicit stack); the substance of the round is the ML oral. For the activation question, the key point is that a purely linear function provides no non-linearity, so stacking layers collapses to a single linear map; a small negative slope is what makes Leaky ReLU useful, not linearity itself. Candidates who have not kept CNN fundamentals fresh get caught here — the interviewer expects textbook ML-course recall delivered verbally and quickly. Preparation Re-derive why convolution gives parameter sharing and translation equivariance versus a fully connected MLP, and be able to state it in two sentences. Explain residual connections and the vanishing-gradient motivation behind ResNet from memory. Drill activation functions (sigmoid / tanh / ReLU / Leaky ReLU): shapes, gradients, dead-neuron and vanishing-gradient trade-offs, and why a linear activation defeats depth.