Do I just have really bad overfitting, or is GradCAM marking the wrong thing, cause why does it have such a big confidence with a bowl that could fit dogs and cats?

What does gradcam look is the difference between the highest , maybe use bone radiosldensity as a benchmark for your calculations

It's my last rodeo on GradCAM with this project. Up next, let's try our hands on Google's MedSigLIP 😄

🌍 Just completed a full End-to-End Image Classification project on Intel's Natural Scenes Dataset! As a Data Scientist & ML Engineer, I built a complete pipeline that automatically identifies 6 types of natural scenes — Buildings, Forest, Glacier, Mountain, Sea & Street — from raw images. 📌 What I did: ✅ Exploratory Data Analysis (EDA) — class distributions, brightness/contrast stats, color histograms ✅ Data Augmentation — rotation, zoom, flip, brightness tuning to prevent overfitting ✅ Built & compared 4 Deep Learning models: 🔹 Simple CNN (Baseline) 🔹 Deep Custom CNN with BatchNorm Dropout 🔹 VGG16 Transfer Learning (freeze fine-tune) 🔹 MobileNetV2 Transfer Learning ✅ Grad-CAM visualizations — to explain WHAT the model actually sees ✅ Confusion Matrix, Classification Report & Per-Class Accuracy ✅ Final predictions exported as CSV with confidence scores 📊 Dataset: ~25,000 images | 6 classes | 150×150 px 🏆 Best Model Accuracy: 94% 💡 Key Takeaway: Transfer Learning is a game-changer. MobileNetV2 & VGG16 significantly outperformed custom CNNs — and Grad-CAM made the model explainable to non-technical stakeholders. 🚀 If your business needs: → Image classification or object detection solutions → Computer Vision pipelines for automation → Explainable AI for stakeholder reporting Let's connect and talk! 📩 DM me or drop a comment below. #MachineLearning #DeepLearning #ComputerVision #CNN #TransferLearning #ImageClassification #VGG16 #MobileNetV2 #GradCAM #Python #TensorFlow #Keras #DataScience #AI #NeuralNetworks #Kaggle #OpenToWork #MLEngineer #DataScientist #AIFreelancer #ClientWork #Portfolio #BuildInPublic #ArtificialIntelligence #MLOps #ExplainableAI #LinkedInLearning #TechPakistan #PakistaniDeveloper #FreelancePakistan

I'm starting 100 days of Deep Learning by CampusX. I learnt so much of dl in the past few days while preparing for an internship. I built EmoVision, a facial emotion recognition CNN on FER-2013 with GradCam, I Implemented ResNet-18 from scratch in PyTorch on CIFAR data, I built CrowdLens, Implementing CSRNet for crowd density estimation using dilated CNNs on ShanghaiTech data. I can't wait to learn more and I hope to actually do some research in dl :)

Day 153 of AI/ML Studies: > done with facial emotion recognition using CNN gradCam on FER-2013 > spent some more time learning the code > continued ResNet-18 from scratch in PyTorch.

How to protect yourself: - Always check what your model actually learned (use SHAP, LIME, GradCAM) - Watch for confounding: are your labels correlated with batch, site, or technical variables?

Get hands-on experience in AI with Explainable AI: Scene Classification and GradCam Visualization. imp.i384100.net/c/6457882/12… #GenAI #AI #AD #Courses

The future is now! Learn about Deep Learning with PyTorch : GradCAM and shape tomorrow. imp.i384100.net/c/6457882/12… #DeepLearning #MachineLearning #AD

Introducing, PulmoScan AI - A high-sensitivity Pneumonia Screening tool powered by Deep Learning > With GradCAM visuals. > DenseNet121 - misses rarely 1 cases with less false alarms > Higher recall ~99.7% Check Out the Live App:pulmoscan.streamlit.app

Introducing, PulmoScan AI - A high-sensitivity Pneumonia Screening tool powered by Deep Learning > With GradCAM visuals. > DenseNet121 - misses rarely 1 cases with less false alarms > Higher recall ~99.7% Vs. ResNet50 ~97% App:pulmoscan.streamlit.app Repo Link in comment.

Introducing, PulmoScan AI - A high-sensitivity Pneumonia Screening tool powered by Deep Learning > With GradCAM visuals. > DenseNet121 - misses rarely 1 cases with less false alarms > Higher recall ~99.7% Vs. ResNet50 ~97% App:pulmoscan.streamlit.app Repo Link in comment.

Now Adding GradCam visuals to the project

【医療×AI】医療AIの「説明」が、素人を誤診に導き、専門家の判断力を高める——同じ技術が真逆の結果を生む衝撃の研究結果 📌 ポイントまとめ ・LLMの説明は一般市民には「諸刃の剣」：AIが正しければ 13%、間違えば-21%の精度変化 ・医療専門家はAIの誤りに惑わされず、自分の判断を維持できる「レジリエンス」を発揮 ・AIを先に見せる「AI-First」より、人間が先に判断する「Human-First」がバイアス軽減に有効 コロンビア大・MIT・スタンフォード大などの共同研究チームが、623名の一般市民と153名の総合診療医を対象に皮膚科診断の実験を実施。GPT-4Vを使ったLLM説明、GradCAM、CBIRなど4種類のXAI（説明可能なAI）を比較しました。 興味深いのは、LLMの「もっともらしい説明」が専門知識のない人ほど過信を招きやすい点。一方で専門家は批判的に評価し、確信度と精度の整合性が向上したとのこと。 今後の医療AI設計では、ユーザーの専門性に応じた「適応型XAI」が鍵になりそうです🔬 興味のある方はぜひチェックしてみてください！ 詳細は👉 scholarcompass.net/articles/… #スカコン #医療AI #LLM #説明可能なAI

Why do you think it’s the case? Is it because researchers learn gradcam in some DL intro classes and are not familiar with anything else? Is it because it’s relatively simple to implement?

@MattNiessner Excited you're heading to Hawaii for #ICCV25! I'll be there, presenting "Sparse Hyperbolic Convolutional Networks with Enhanced Object Localisation via GradCAM Analysis" at the BEW workshop. Would love to meet!

best Ai tool for building gradcam??

3D CNNと転移学習を用いたParkinson病の診断/進行予測 system PD診断で74.04%の診断制度(イマイチか) 前処理最小化と多施設間汎用性により実臨床実装が可能でgradCAMによる判断根拠可視化も実現 OA #parkinsons #CNN #papers doi.org/10.1016/j.nicl.2025.…

we are researching gradcam too xD colab.research.google.com/dr…