Building Your Own Classification Model with JavaScript - Coffee with Developers - Carly Richmond

What if your machine learning model to identify cake… couldn't identify any cake at all? Learn how one developer solved this problem using JavaScript.

#1about 3 minutes

Building a machine learning game with JavaScript

A side project inspired by the Netflix show "Is It Cake?" was created to experiment with machine learning in JavaScript using TensorFlow.js.

#2about 2 minutes

Training a custom binary classifier from scratch

The process of building a custom binary classifier involved collecting a dataset of images and training the model to extract features without overfitting.

#3about 1 minute

Diagnosing why the custom classification model failed

The custom model failed to identify cakes, likely due to insufficient training data, the use of color images instead of monochromatic ones, or simple coding errors.

#4about 3 minutes

Improving accuracy with transfer classification and MobileNet

Using transfer classification with the pre-trained MobileNet model significantly improved cake detection accuracy compared to building a model from scratch.

#5about 3 minutes

The developer trend of consuming vs creating AI

Modern developers often prefer consuming off-the-shelf AI solutions due to system complexity and time pressures, which can reduce natural curiosity and deep learning.

#6about 4 minutes

How engineering managers can stay technically hands-on

Managers can stay technical by purposefully blocking out calendar time for coding and seeking roles or company cultures that expect and support their hands-on contributions.

#7about 5 minutes

Fostering innovation with internal hackathons and tinker time

Company-sponsored hackathons or innovation weeks can drive creativity, but they often fail when participation becomes optional under deadline pressure or when winning projects are not implemented.

#8about 3 minutes

The problem of trust and transparency in AI models

While platforms provide model cards to explain training data and limitations, they often lack sufficient detail, making it difficult to assess potential biases and build trust.

#9about 3 minutes

Why the marketplace for custom AI datasets failed

The concept of a marketplace for custom datasets has not taken off due to the dominance of large AI providers, the high cost of training, and a developer preference for established solutions.

#10about 5 minutes

Balancing AI automation and authenticity in content creation

While AI tools can automate content generation, they risk losing the author's authentic voice, shifting the creator's role from writing to validating and editing AI output.

#11about 5 minutes

Addressing the societal risks of deepfakes and misinformation

The ability to generate realistic deepfakes poses a significant risk for public figures and society, highlighting the urgent need for established ethical guidelines on AI usage.